QUT and Baker Lab Turn AI-Designed Proteins Into Molecular Switches That Work Inside Living Cells
QUT and Baker-lab researchers built AI-designed allosteric switches that turn on in the presence of small molecules, peptides, or whole proteins and work inside bacteria or on electrodes.
Overview
An international research team has built a general-purpose method for turning AI-designed proteins into molecular switches whose activity turns on only in the presence of a chosen target. The work, led by Professor Kirill Alexandrov at the Queensland University of Technology (QUT) together with collaborators including 2024 Nobel laureate Professor David Baker at the University of Washington, was published on April 15 in Nature Biotechnology under the title “Artificial allosteric protein switches with machine-learning-designed receptors”.
The result extends a decade of de novo protein design past static binders and into fully functional biosensor components that can operate inside living bacteria or transduce a signal onto an electrode.
What We Know
The paper describes a pipeline that couples machine-learning-designed receptors to a protein scaffold whose activity is switched by the binding event. The researchers built working switches across three classes of targets — small molecules, peptides, and whole proteins — and used the approach to construct electrochemical biosensors for steroid detection, according to Nature Biotechnology.
The QUT-led team included Dr. Zhong Guo, Dr. Zhenling Cui, Dr. Cagla Ergun Ayva, Dr. Roxane Mutschler, and Dr. Mica Fiorito, with collaborators drawn from seven groups across Australia, the United Kingdom, and the United States, including Baker’s lab at the University of Washington and Australia’s national science agency CSIRO, according to a QUT announcement carried by EurekAlert.
A key technical claim in the paper is that effective switching does not require the large conformational rearrangements that have long been assumed necessary for allosteric behaviour. Instead, “binding of the target molecule subtly changes how the protein moves” to activate function, as reported by Phys.org. That observation — that modest shifts in dynamics are enough to drive a measurable output — broadens the design space that engineers can target with AI tools.
The switches were also shown to function inside living bacterial cells and to drive readouts including colour changes, light emission, and electrical signals similar in principle to the glucose meters used by people with diabetes, according to the QUT release. Alexandrov said “AI-designed proteins can be turned into effective molecular switches, greatly expanding what protein engineers can build,” as quoted by EurekAlert.
The work builds on the broader wave of AI-driven protein design that has accelerated since Baker’s 2024 Nobel Prize for computational protein design, and follows The Machine Herald’s earlier reporting on MIT’s VibeGen model, which takes a complementary approach by designing proteins around their vibrational dynamics rather than their static structure.
What We Don’t Know
The Nature Biotechnology paper and accompanying press materials do not disclose quantitative limits of detection, dynamic range, or cross-reactivity for the demonstrated electrochemical steroid biosensors, and neither the Phys.org summary nor the EurekAlert release provides specific efficiency or performance metrics. Those characteristics will determine whether the switches can compete with established antibody- or aptamer-based biosensors in regulated clinical or environmental settings.
It is also unclear how the receptor-design pipeline will generalise beyond the target classes reported. Small molecule recognition has historically been a harder problem for de novo design than protein-protein interfaces, and the authors have not released the full list of analytes they attempted or the hit rate from their machine-learning pipeline.
The long path from a bacterial proof-of-concept to deployed diagnostic or environmental devices also depends on manufacturing, stability under field conditions, and regulatory validation — none of which are addressed in the current publication.
Why It Matters
Most existing biosensors rely on naturally occurring recognition elements — antibodies, aptamers, or enzymes evolved to bind a narrow slice of chemistry. Designing a new sensor typically means either screening vast libraries for an accidentally suitable binder or re-engineering an existing one.
A machine-learning pipeline that produces switchable receptors on demand, for target classes as different as small molecules and whole proteins, would change that economics. It would also move a significant fraction of sensor development from wet-lab screening to computational design, a shift that parallels the one already underway in structural biology since AlphaFold.
The QUT paper stops short of claiming a general solution — it is a demonstration on a chosen set of targets — but by showing that AI-designed receptors can be fused into functional allosteric switches that work in cells and on electrodes, it narrows the gap between de novo protein design and the kinds of point-of-care, environmental, and synthetic-biology applications that such tools have long been promised to enable.